This landmark Stanford GSB study demonstrates that an AI analyst using only publicly available information outperformed 93% of mutual fund managers over a 30-year period (1990-2020). The AI-adjusted portfolios generated $17.1 million in additional quarterly returns compared to human managers' $2.8 million—a stunning ~600% improvement.
The research reveals a critical insight about information costs: the AI's advantage comes not from access to private information, but from superior processing of public data. Counterintuitively, the model primarily relied on simple variables like firm size and trading volume, but deployed sophisticated machine learning techniques to extract maximum predictive value. This demonstrates that "public" information still carries substantial hidden value—a "shadow price"—when processed efficiently.
Testing across 3,300 diversified U.S. equity mutual funds with 170 publicly available variables, the study quantifies what researchers term "information processing costs"—the expensive computation required to fully exploit freely available datasets. This has profound implications for AI-augmented investment management.
Imagine a library where all books are free to read. A human librarian can read maybe 10 books and remember key points. But an AI "librarian" can read every book, remember every detail, and spot patterns humans miss—like noticing that authors who use certain words tend to write bestsellers. The books are free, but the ability to read them all and find hidden connections? That's incredibly valuable. This study proves that in investing, the "free" public data is worth billions when you can actually process it all.
Professional fund managers have long justified their fees by claiming access to superior information and analysis. But in an era of mandatory disclosure requirements, most market-relevant data is technically "public." The question becomes: if everyone has access to the same data, why do some managers consistently outperform?
The researchers introduce the concept of "shadow price"—the true cost of extracting value from public information. While datasets themselves may be freely available, the computational resources and analytical sophistication required to fully exploit them create substantial barriers.
This explains why active management persists despite efficient market theory—the "efficiency" assumes perfect information processing, which humans cannot achieve.
The researchers employed a random forest model for predicting quarterly benchmark-adjusted returns:
The AI analyzed 170 publicly available variables, including:
| Category | Example Variables | Source |
|---|---|---|
| Market Data | Firm size, dollar trading volume, price momentum | Stock exchanges |
| Fundamentals | P/E ratios, book value, debt levels | SEC filings (10-K, 10-Q) |
| Fixed Income | Treasury rates, credit spreads, yield curves | Federal Reserve |
| Sentiment | Earnings call tone, regulatory filing sentiment | NLP analysis of transcripts |
| Analyst Data | Ratings changes, price targets, EPS estimates | Sell-side research |
| Macro | GDP growth, inflation, unemployment | Government agencies |
Counterintuitively, the AI primarily relied on the simplest variables—firm size and dollar trading volume—rather than sophisticated metrics. The advantage came from deploying "complex AI techniques to squeeze the most predictive value from this simple data." This suggests that expertise lies not in finding obscure signals, but in optimally processing obvious ones.
The results demonstrate a dramatic performance gap between AI-assisted and human-only management:
| Metric | Human Managers | AI-Adjusted | Improvement |
|---|---|---|---|
| Quarterly Alpha | $2.8 million | $17.1 million | +$14.3M (+511%) |
| Manager Ranking | Median performance | Top 7% | Beat 93% of managers |
| 30-Year Performance | Baseline | +600% vs human average | 6× better returns |
| Holdings Modified | N/A | ~50% quarterly | Active rebalancing |
The AI rebalanced holdings quarterly without fundamentally altering fund characteristics like risk levels or stock count. It sorted investment options into ten performance-expectation buckets, swapping underperformers for similar assets with better prospects while maintaining portfolio structure. This approach modified approximately 50% of total holdings each quarter, generating sixfold returns over the 30-year simulation.
Even more striking: AI-only portfolios allocated 42% to passive index funds yet still dramatically outperformed active managers. This suggests that a substantial portion of active management fees are wasted on decisions that underperform simple indices.
The findings suggest investment firms should automate data collection and analysis work currently performed by analysts. However, co-author Ed deHaan notes: "There will always be a role for clever humans who can guide the process and think in broad ways about strategies." The optimal model appears to be AI-human collaboration, not replacement.
This study provides crucial context for understanding why LLM-based trading agents (evaluated in StockBench and AI-Trader) struggle as direct traders but show promise as analysts:
| Approach | Task Type | Performance | Source |
|---|---|---|---|
| AI as Analyst | Information processing | +600% vs humans | This paper |
| LLM as Trader | Real-time decisions | Most fail vs buy-and-hold | StockBench, AI-Trader |
| Multi-Agent System | Collaborative analysis | Outperforms single-agent | AlphaAgents |
The emerging picture suggests a clear division: AI excels at processing vast public information and identifying statistical patterns, but struggles with real-time market dynamics and regime recognition. The optimal architecture may be AI-generated signals feeding into human (or more sophisticated agent) decision-making—not end-to-end automation.
The researchers acknowledge significant limitations:
These limitations mirror the "alpha decay" phenomenon in quantitative finance—strategies that work in backtests often fail once widely adopted.
This Stanford GSB research fundamentally reframes the value proposition of AI in finance. The key insight is not that AI has access to better information, but that it can extract more value from existing public information than humans possibly can.
For AI agent engineering, this study validates the approach of using AI for information processing and signal generation, while reserving dynamic decision-making for architectures better suited to real-time adaptation.
The Shadow Value of "Public" Information: Evidence from Mutual Fund Managers
deHaan, Lee, Liu, Noh — Stanford GSB Research Paper, July 2025
An AI Analyst Made 30 Years of Stock Picks — and Blew Human Investors Away
Stanford Graduate School of Business Insights
Ed deHaan Faculty Page
Stanford GSB Professor of Accounting